Sleep Quality Evaluation Based on Single-Lead Wearable Cardiac Cycle Acquisition Device

Chengyu Liu, Yang Li, Jianqing Li, Chang Yan, Kejun Dong, Zhiyu Kang, Hongxing Zhang, Chengyu Liu 0001Sensors 2023 · DOI

TL;DR

This paper introduces a novel, two-step method to evaluate sleep quality using only a single-lead wearable ECG device. By first classifying sleep stages and then using those features to predict subjective sleep quality, the approach offers a practical, low-cost alternative to traditional polysomnography for daily sleep health monitoring.

What Problem Does This Paper Address?

Accurate sleep quality assessment is crucial for both clinical diagnosis and general well-being. The current gold standard, polysomnography (PSG), is cumbersome, expensive, and can interfere with natural sleep due to its numerous sensors and wires. While sleep questionnaires offer a subjective measure, they can be unreliable. There's a clear need for a simpler, more accessible, and objective method to evaluate sleep quality, especially for everyday monitoring and for individuals in demanding professions (e.g., soldiers, firefighters) who require optimal sleep. This paper aims to bridge this gap by leveraging readily available single-lead ECG data.

Methods

The proposed method employs a two-stage approach:

  1. Sleep Staging Model: This model classifies sleep into four stages: wakefulness (W), REM, light sleep (N1&N2), and deep sleep (N3). It utilizes a multi-classifier fusion method.
  2. Feature Extraction: Heart Rate Variability (HRV) features and ECG-derived respiration features are extracted from the ECG signals. A comprehensive set of 23 candidate features, including linear HRV metrics, nonlinear features, and time irreversibility features, were initially considered.
  3. Feature Selection: A rigorous feature selection process was employed, including variance analysis and Support Vector Machine Recursive Feature Elimination (SVM-RFE), to identify the most discriminative features for each sleep stage.
  4. Model Construction: A novel approach involves constructing individual "recognizers" for each sleep stage. These recognizers use the first principal component of selected feature combinations and a variable threshold with window adjustment to determine sleep stage transitions. A multi-classifier fusion method, incorporating sigmoid fitting and conditional probabilities, is then used to combine the outputs of these recognizers into a final sleep stage classification.
  5. Sleep Quality Evaluation Model: This model takes the output from the sleep staging model and uses it to predict subjective sleep quality, categorizing it as "good sleep" or "poor sleep."
    • Feature Extraction: Features reflecting sleep fragmentation (e.g., occurrence of successive W stages, sleep stage transition intensity), sleep depth, and sleep duration are derived from the sleep staging results.
  6. Model Construction: A Support Vector Machine (SVM) model, with a third-order polynomial kernel, is trained using these extracted features and the subjective sleep questionnaire scores as ground truth. SVM-RFE is again used for feature selection.

Datasets

The study utilized two primary international public datasets for training and validating the sleep staging model:

  • Sleep Heart Health Study Visit 1 (SHHS1DB): This dataset contains PSG data from 6441 individuals.
  • PhysioNet Computing in Cardiology Challenge 2018 training database (CincDB): This dataset includes PSG data from 994 subjects, some of whom were diagnosed with sleep disorders.

For the sleep quality evaluation model, the researchers collected ECG data and sleep questionnaires from 200 healthy male subjects working in special positions. This dataset was split into training and testing sets.

Key Insights and Ideas

  • Single-lead ECG is Sufficient: The study demonstrates that a single-lead ECG, a readily available signal from wearable devices, can provide enough information for both sleep staging and sleep quality evaluation, significantly simplifying the process compared to multi-lead PSG.
  • Bridging Objective and Subjective: The two-step approach effectively bridges the gap between objective physiological signals (ECG) and subjective self-reported sleep quality. The sleep staging acts as a crucial intermediary, translating raw ECG data into meaningful features for quality assessment.
  • Novel Sleep Staging Recognizers: The development of individual "recognizers" for each sleep stage, utilizing principal components and adaptive thresholds, represents a novel approach to sleep stage classification from ECG.
  • Feature Importance for Sleep Stages: The research highlights that different feature combinations are optimal for recognizing different sleep stages, underscoring the complexity of sleep stage classification and the need for stage-specific feature sets.
  • Practicality and Portability: The emphasis on using minimal features and computationally efficient methods makes the proposed model highly transplantable to portable devices, facilitating widespread adoption for daily sleep monitoring.

Key Results

  • Sleep Staging Performance: The sleep staging model achieved an accuracy of 0.661 and an average F₁-score of 0.625 on the CincDB, and 0.659 accuracy and 0.624 F₁-score on the SHHS1DB.
  • Sleep Quality Evaluation Performance: The sleep quality evaluation model demonstrated an accuracy of 0.786 and an average F₁-score of 0.771 on the test set, using subjective questionnaire scores as the ground truth.
  • Feature Reduction: The feature selection process successfully reduced the number of features required for the models, enhancing their efficiency and suitability for portable devices.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Gender Differences: The current study primarily included male subjects, and future work should incorporate more female participants to explore potential gender differences in sleep patterns and model performance.
  • Sleep Apnea Exclusion: The dataset excluded individuals with sleep apnea. Future research should investigate the model's performance on patients with sleep disorders.
  • Night-to-Night Variability: The study did not extensively consider night-to-night variability in sleep patterns, which could be an area for improvement.
  • ECG-based Sleep Staging Precision: While promising, ECG-based sleep staging still lags behind EEG in precision. Further research is needed to improve the accuracy of ECG-based sleep stage discrimination.
  • Real-World Data and Cross-Dataset Validation: More extensive testing with real-world data and robust cross-dataset validation are necessary to improve the model's generalization ability.
  • Relationship between Sleep Stages and Quality: The lack of precise sleep stage labeling for the special position staff dataset limits the ability to definitively establish the exact relationship between specific sleep stages and overall sleep quality.
  • Adaptability to Different Populations/Diseases: The algorithm may require modification and improvement to be applicable to diverse populations and different medical conditions.
  • Optimization: Future work will focus on optimizing and improving the algorithm to suit various age groups, populations, and sleep disorders.

Summary updated

BibTeX
@article{liu2023sleep,
  author = {Chengyu Liu and Yang Li and Jianqing Li and Chang Yan and Kejun Dong and Zhiyu Kang and Hongxing Zhang and Chengyu Liu},
  title = {{Sleep Quality Evaluation Based on Single-Lead Wearable Cardiac Cycle Acquisition Device}},
  journal = {Sensors},
  year = {2023},
  doi = {10.3390/s23010328},
}
Figure